Fact-Checking with Large Language Models via Probabilistic Certainty and Consistency

TL;DR

This paper introduces PCC, a framework that improves fact-checking in large language models by estimating their confidence through probabilistic certainty and consistency, enabling adaptive retrieval and enhancing accuracy and efficiency.

Contribution

The paper presents PCC, a novel confidence estimation method that guides adaptive fact-checking in LLMs, outperforming existing baselines and generalizing across models.

Findings

PCC achieves superior uncertainty quantification compared to verbalized confidence.

PCC outperforms strong LLM-based fact-checking baselines across benchmarks.

PCC generalizes well across various large language models.

Abstract

Large language models (LLMs) are increasingly used in applications requiring factual accuracy, yet their outputs often contain hallucinated responses. While fact-checking can mitigate these errors, existing methods typically retrieve external evidence indiscriminately, overlooking the model's internal knowledge and potentially introducing irrelevant noise. Moreover, current systems lack targeted mechanisms to resolve specific uncertainties in the model's reasoning. Inspired by how humans fact-check, we argue that LLMs should adaptively decide whether to rely on internal knowledge or initiate retrieval based on their confidence in a given claim. We introduce Probabilistic Certainty and Consistency (PCC), a framework that estimates factual confidence by jointly modeling an LLM's probabilistic certainty and reasoning consistency. These confidence signals enable an adaptive verification strategy: the model answers directly when confident, triggers targeted retrieval when uncertain or inconsistent, and escalates to deep search when ambiguity is high. Our confidence-guided routing mechanism ensures that retrieval is invoked only when necessary, improving both efficiency and reliability. Extensive experiments across three challenging benchmarks show that PCC achieves better uncertainty quantification than verbalized confidence and consistently outperforms strong LLM-based fact-checking baselines. Furthermore, we demonstrate that PCC generalizes well across various LLMs.